IRG XGBoost baselines (notebook 09)
Fifteen gradient-boosted classifiers (3 biosamples x 5 seeds) that predict whether a 1,000 bp window is an ENCODE candidate
cis-regulatory element (cCRE: PLS, pELS or dELS) of one biosample, against GC-matched genomic background, from 6-mer
frequencies (4,096 features). Trained on PxALabs/encode-regulatory-elements v2.1, *_gc_matched tables. Part of
github.com/PxA-Labs/interpretable-regulatory-genomics. No pretrained or language models are involved. These are the tuned
reference baseline that CNNs and any later model must be compared against.
Files
weights/<biosample>/seed<N>.ubj- XGBoost native binary models (tcell,k562,gm12878; seeds 0-4); no picklexgb_features.py- the feature function (6-mer frequencies; windows containing N are skipped)manifest.json- per file: SHA-256, size, selected hyperparameters, best iteration, validation and test AUROC;MANIFEST.sha256
Usage
import xgboost as xgb
from xgb_features import kmer_features
clf = xgb.XGBClassifier(); clf.load_model("weights/k562/seed0.ubj")
p = clf.predict_proba(kmer_features(["ACGT" * 250], k=6))[:, 1] # probability of being a K562 cCRE
Use all five seeds and report the spread; do not pick a seed on test data.
Training
Random search over 24 configurations on the validation chromosomes only (seed 0, early stopping on validation AUC, up to
1,500 trees, patience 50). Search space: max_depth {4, 6, 8}, learning_rate {0.05, 0.1}, subsample {0.6, 0.8, 1.0},
colsample_bytree {0.3, 0.5, 0.8}, min_child_weight {1, 5, 10}, reg_lambda {1, 5, 20}; configuration 0 is the earlier default.
Selected: tcell (depth 6, lr 0.05, subsample 0.6, colsample 0.3, min_child_weight 10, lambda 1); k562 (depth 6, lr 0.05, subsample 1.0,
colsample 0.3, min_child_weight 5, lambda 20); gm12878 (depth 8, lr 0.05, subsample 1.0, colsample 0.3, min_child_weight 1, lambda 20).
Split: train chr1-14, X, Y / validation chr15-18 / test chr19-22. Test chromosomes were never used for selection.
Source: Kaggle kernel purvanshj009/irg-09-xgb-tuned v2, git ff83f8f, XGBoost on GPU.
Evaluation (test chromosomes, mean over 5 seeds; SD in brackets)
"Matched" = test subset matched on GC and CpG observed/expected, equal positives and negatives per bin (a sub-population; not comparable with the full-task AUROC).
| biosample | test AUROC | matched AUROC | untuned default (same code) |
|---|---|---|---|
| tcell | 0.8894 (0.0003) | 0.7431 (0.0012) | 0.8845 / 0.7347 |
| k562 | 0.8695 (0.0004) | 0.7686 (0.0011) | 0.8636 / 0.7614 |
| gm12878 | 0.8429 (0.0008) | 0.7526 (0.0011) | 0.8350 / 0.7452 |
Against the tuned CNNs of PxALabs/irg-cnn-baselines, these models are level on T cell and K562 and slightly behind on GM12878
(ShallowCNN +0.004 full, +0.012 matched). Full analysis and caveats: docs/results/xgb_tuned.md in the GitHub repository.
Intended use and limitations
- Reference baselines for sequence-based cCRE classification. Not for clinical use or variant-effect prediction.
- Each model is specific to the biosample it was trained on; cross-biosample use is not covered here.
- The negatives are GC-matched genomic windows outside the biosample's cCREs, not validated non-regulatory elements.
- The search was truncated. The selected configurations sit on the edge of the search space (lowest learning rate and colsample tried; highest lambda in two biosamples), and the best iteration was 1,369-1,499 of the 1,500-tree cap, so these models are probably still slightly under-tuned.
- Only k=6 was tuned. The 24-configuration search used one seed, and validation differences between configurations were small (about 0.005-0.007).
- Composition shortcuts (GC, CpG) are controlled only in the matched evaluation; repeat content is not controlled in training.
- The untuned-default comparison models were not saved.
- Data note: the project's earlier tables named
k562_*were T-cell elements; these models use the corrected v2.1 tables.
Provenance and integrity
Every file was reloaded and its best iteration checked against the run log. Seed-0 models for each biosample and seed 3 for GM12878 were re-scored on the real test rows and reproduced the AUROC recorded during the run to four decimals (0.8895, 0.8698, 0.8419, 0.8436). The other 11 were checked for loading and best iteration only.